According to the arXiv abstract, recent literature has shown that language models can issue textual refusals while still performing unsafe tool calls. This discrepancy represents a hidden vulnerability that standard safety evaluations might miss.

To investigate this problem, the authors introduce PRISM, a framework designed to evaluate the transfer of activation interventions in tool-calling language models. The abstract indicates that the framework is aimed at understanding how such interventions behave when applied to tool-calling scenarios.

The paper is a single preprint, so its claims have not yet been independently verified. Still, the framing suggests a growing concern with safety gaps between what a model says and what it actually does when given tools.